Perceptive Humanoid Parkour (PHP) framework enables real-time vision-based parkour on Unitree G1 via motion matching

The Perceptive Humanoid Parkour framework is a modular system that enables humanoid robots to perform long-horizon, vision-based parkour across obstacle courses. The approach first retargets human motion data into atomic skills using OmniRetarget. It then applies motion matching as nearest-neighbor search in a feature space to chain these skills into diverse kinematic trajectories that preserve fluid human motion characteristics. Motion-tracking reinforcement learning expert policies are trained on the composed trajectories. These policies are distilled into a single depth-conditioned student policy through a combination of DAgger and reinforcement learning. The resulting policy allows a robot to select and execute behaviors such as stepping over, climbing onto, vaulting, or rolling off obstacles using only onboard depth sensing and a discrete two-dimensional velocity command. Researchers validated the framework through real-world tests on a Unitree G1 humanoid robot. The system achieved climbing of obstacles up to 1.25 meters, or 96 percent of the robot's height, along with multi-obstacle courses that include closed-loop adaptation to real-time perturbations. The work demonstrates zero-shot sim-to-real transfer of the depth-based policy.